A multivariate method to determine the dimensionality of neural representation from population activity

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Neuronal populations encode variables through activation patterns in feature spaces.
• The dimensionality of these feature spaces is crucial for understanding neural representations.
• Existing methods for determining dimensionality are limited.

Purpose of the Study:

• To introduce a novel method for determining the number of feature dimensions in neuronal population codes.
• To provide a tool for quantitative modeling of neuronal activity.
• To analyze the dimensionality of motor cortical representations of force and finger presses.

Main Methods:

• Proposed a method using Gaussian linear classifiers on principal components of activation patterns.
• Employed cross-validation to select the optimal dimensionality.
• Applied the method to functional magnetic resonance imaging (fMRI) data of motor cortex activity.

Main Results:

• The representation of force levels was found to be low-dimensional, with scaled activation patterns.
• Individual finger presses were represented in a higher, four-dimensional feature space.
• The method successfully identified the dimensionality without prior knowledge of the features.

Conclusions:

• The new method effectively determines the dimensionality of neuronal population codes.
• Neuronal representations of force and individual finger movements differ significantly in dimensionality.
• This approach offers a valuable tool for analyzing neural population activity and building quantitative models.